WhatsGNU-ATB

A custom reimplementation of WhatsGNU optimised for the scale of AllTheBacteria. It uses LMDB-backed sharded storage (8 shards) with numpy for hashing. The query tool is also custom-built for this database format. Protein allele frequency analysis at the scale of AllTheBacteria (2.4M+ bacterial genomes).

WhatsGNU-ATB builds a sharded LMDB database from Bakta protein annotations and lets you query any bacterial genome to find out, for each protein, how many of the 2,438,285 genomes carry an identical copy — along with which species they belong to and which genomes are most similar.

A pre-built database covering all AllTheBacteria genomes is available on OSF. If you just want to query genomes, skip to Quick Start (Query).

Features

• GNU scores: for every protein in a query genome, reports the exact number of genomes (out of 2.4M+) containing an identical allele
• Species breakdown: top-K species contributing to each allele, with counts (other metadata like MLST contributions are coming soon)
• Genome similarity: ranks all 2.4M+ genomes by shared protein alleles with your query, identifying the closest relatives
• Batch querying: pass a directory of .faa files to query hundreds of genomes in one run
• Sequence export: optionally include the amino acid sequence in the output
• Sharded LMDB backend: 8 parallel shards with batched reads for fast lookups
• Optional sequence storage: store a representative amino acid sequence per allele hash in the database
• Allele counts export: dump the full allele frequency table as a TSV

Installation

Option A — Conda (recommended, once available on bioconda)

conda install -c bioconda whatsgnu-atb

Option B — pip

pip install whatsgnu-atb

Option C — From source

git clone https://github.com/microbialARC/WhatsGNU-ATB.git
cd WhatsGNU-ATB
bash setup_whatsgnu_atb.sh
conda activate whatsgnu-atb

Option D — Manual from source

conda create -n whatsgnu-atb -c conda-forge python=3.12
conda activate whatsgnu-atb
pip install numpy lmdb pandas

git clone https://github.com/microbialARC/WhatsGNU-ATB.git

For publication figure generation, also install:

pip install matplotlib seaborn networkx adjustText scipy

Quick Start (Query)

If you just want to query genomes against the pre-built AllTheBacteria database:

1. Download the database from OSF

Use the included downloader (no OSF account or token required):

# Download the database (required for querying)
python scripts/download_osf.py --folder WGNU_ATB_DB --out-dir ./WGNU_ATB_DB

# Download everything
python scripts/download_osf.py --all --out-dir ./whatsgnu_db

# List available folders
python scripts/download_osf.py --list

The downloader skips files that have already been downloaded with the correct size, so it is safe to rerun if interrupted.

2. Query a single genome

Your input must be a protein FASTA (.faa) file. See the AllTheBacteria Bakta documentation or the Bakta GitHub if you need to annotate your genome first.

Basic query (GNU scores only — fast, no postings needed):

python scripts/Query_WhatsGNU_ATB.py \
    --db_dir WGNU_ATB_DB/ \
    --shards 8 \
    --faa your_genome.bakta.faa \
    --out_dir results/

Full query (GNU scores + species breakdown + genome similarity):

python scripts/Query_WhatsGNU_ATB.py \
    --db_dir WGNU_ATB_DB/ \
    --shards 8 \
    --faa your_genome.bakta.faa \
    --include_sequence \
    --with_postings \
    --samples_tsv WGNU_ATB_DB/samples_with_ids.tsv \
    --species_names_tsv WGNU_ATB_DB/samples_with_ids.tsv \
    --top_k_species 5 \
    --top_k_genomes 10 \
    --out_dir results/

3. Query a batch of genomes

Pass a directory instead of a single file:

python scripts/Query_WhatsGNU_ATB.py \
    --db_dir WGNU_ATB_DB/ \
    --shards 8 \
    --faa directory_of_faa_files/ \
    --include_sequence \
    --with_postings \
    --out_dir results_batch/

Note: If you installed via conda or pip, the scripts are on your PATH and you can run Query_WhatsGNU_ATB.py, WhatsGNU_ATB_DB.py, and download_osf.py directly without the scripts/ prefix.

OSF Data

All data is hosted at https://osf.io/6jr4u/:

Folder	Description
WGNU_ATB_DB/	Pre-built LMDB database (8 count + 8 posting shards, genome-to-species index, function lookup table, Sample-to-ID mapping (samples_with_ids.tsv), build metadata). Required for querying.
Sample_tables/	List of included genomes (final_2438285_genomes.txt), species statistics, and per-genome/per-species allele record counts.
ATB_hash_seq/	Hash-to-amino-acid-sequence lookup table, split into 20 xz-compressed parts (hash_to_sequence_part_00.xz – part_19.xz).
ATB_summary_figures_tables/	Publication figures, per-species GNU histograms, allele frequency tables, species-sharing networks, coverage estimates, cross-species allele analyses, and the pre-computed counts cache.

Query Output Files

<sample>.whatsgnu.tsv

Per-protein results with one row per protein:

Column	Description
protein_id	Protein identifier from the FASTA header
allele_hash	128-bit BLAKE2b hash of the amino acid sequence (hex)
sequence	Amino acid sequence from the query genome (if --include_sequence)
GNU_count	Number of genomes containing this exact allele
top5_species_names	Top 5 species carrying this allele (if --with_postings)
top5_species_counts	Counts per species (if --with_postings)
total_db_hits	Total genomes in posting list
hits_checked	Number of postings actually decoded

<sample>.similarity.tsv

Genome similarity ranking (if --with_postings):

Column	Description
rank	Rank by shared alleles (1 = most similar)
genome_id	Integer genome ID
sample_name	Sample accession (if --samples_tsv provided)
species_id	Species integer ID
species_name	Species name (if --species_names_tsv provided)
shared_alleles	Number of identical proteins shared with query
percent_of_query	Shared alleles as percentage of query proteome

Query Options Reference

Option	Description	Default
--db_dir	Root database directory (required)	—
--shards	Number of shards, must be power of 2 (required)	—
--faa	Input .faa file or directory of .faa files (required)	—
--out_dir	Output directory (required)	—
--with_postings	Enable species breakdown and genome similarity	off
--include_sequence	Include amino acid sequence in output	off
--top_k_species	Number of top species to report per protein	5
--top_k_genomes	Number of top similar genomes to report	10
--postings_limit	Max genome IDs to decode per allele (0 = all)	0
--species_names_tsv	TSV mapping SpeciesID → species name	none
--samples_tsv	TSV mapping SampleID → sample accession	none

Interpreting GNU Scores

GNU Score Range	Interpretation
>100,000	Highly conserved ubiquitous allele
1000–10,000	Common allele
1–100	Rare allele, likely strain-specific
0	Unique to the query genome — not in any AllTheBacteria genome

Building a Database

To build a new database from scratch (e.g., for a custom genome set):

Input Requirements

A sample table TSV with these columns:

Column	Description
SampleID	Unique integer ID per genome
Sample	Sample name (used to find .faa file)
SpeciesID	Integer species ID

Optional column: faa_path (full path to FAA file). If absent, uses --faa_dir/<Sample><faa_suffix>.

Build Command

python scripts/WhatsGNU_ATB_DB.py \
    --sample_table samples_with_ids.tsv \
    --faa_dir /path/to/faa_files/ \
    --out_dir WGNU_ATB_DB/ \
    --tmp_dir /scratch/tmp/ \
    --shards 8 \
    --with_postings \
    --sort_mem_mb 65536 \
    --lmdb_map_gb_counts_per_shard 24 \
    --lmdb_map_gb_postings_per_shard 160 \
    --export_allele_counts allele_counts.tsv \
    --log_file build.log \
    --log_level INFO

Build with Sequences

To also store representative amino acid sequences per allele hash:

python scripts/WhatsGNU_ATB_DB.py \
    --sample_table samples_with_ids.tsv \
    --faa_dir /path/to/faa_files/ \
    --out_dir WGNU_ATB_DB/ \
    --shards 8 \
    --with_postings \
    --with_sequences \
    --lmdb_map_gb_sequences_per_shard 25 \
    --log_level INFO

Build Options Reference

Option	Description	Default
--sample_table	Sample table TSV (required)	—
--faa_dir	Directory of .faa files	none
--out_dir	Output directory (required)	—
--tmp_dir	Temp directory for intermediate files	<out_dir>/tmp
--reduce_tmp_dir	Local scratch for sort/reduce (faster I/O)	none
--shards	Number of shards, power of 2	16
--with_postings	Build posting lists (genome IDs per allele)	off
--with_sequences	Store representative AA sequence per allele	off
--faa_suffix	Suffix appended to Sample name for FAA lookup	.bakta.faa
--sort_mem_mb	RAM for external sort per shard (MB)	65536
--lmdb_map_gb_counts_per_shard	LMDB map size for counts (GB)	24
--lmdb_map_gb_postings_per_shard	LMDB map size for postings (GB)	160
--lmdb_map_gb_sequences_per_shard	LMDB map size for sequences (GB)	25
--export_allele_counts	Path to write allele frequency TSV	none
--parse_only	Only parse FAA → record bins, skip reduce	off
--reduce_only	Only reduce existing record bins → LMDB	off
--resume	Auto-detect: skip parse if record bins exist	off
--skip_existing_shards	Skip shards with existing LMDB output	off
--log_file	Log file path	<out_dir>/build.log
--log_level	Logging level	INFO

Database Structure

WGNU_ATB_DB/
├── lmdb_counts/
│   ├── shard_00/         # LMDB: hash → (func_id, GNU_count)
│   ├── shard_01/
│   └── ...
├── lmdb_postings/        # (if --with_postings)
│   ├── shard_00/         # LMDB: hash → varint-encoded genome IDs
│   ├── shard_01/
│   └── ...
├── lmdb_sequences/       # (if --with_sequences)
│   ├── shard_00/         # LMDB: hash → amino acid sequence (UTF-8)
│   └── ...
├── indexes/
│   └── genome_species.u32   # Binary array: genome_id → species_id
└── metadata/
    ├── build_info.json       # Build parameters, stats, version
    └── functions.tsv.gz      # Function ID → function description

Technical Details

• Hashing: BLAKE2b with 128-bit (16-byte) digest of the amino acid sequence
• Sharding: shard_id = first_byte(hash) & (num_shards - 1)
• GNU count: number of genomes containing an allele at least once (deduplicated within each genome)
• Postings: delta + varint encoded sorted unique genome IDs
• External sort: numpy structured arrays for memory-efficient sorting; batched multi-pass merge with fanin of 64
• Query optimizations: batched LMDB reads (one transaction per shard), numpy-vectorized species lookups, partial argsort for top-K genome ranking

Resource Requirements

Building (2.4M genomes, 8 shards)

Resource	Recommendation
RAM	250–500 GB
CPUs	4–6 cores
Disk (tmp)	~2 TB scratch
Wall time	6–24 hours (I/O dependent)

Querying

Resource	Recommendation
RAM	~2 GB (basic) / ~4 GB (with postings)
Wall time	~5–150 seconds per genome

Citation

If you use WhatsGNU-ATB in your research, please cite:

Moustafa AM and Planet PJ. WhatsGNU: a tool for identifying proteomic novelty. Genome Biology, 2020. doi:10.1186/s13059-020-01965-w

Hunt M, Lima L, Shen W, Lees J, Iqbal Z. AllTheBacteria - all bacterial genomes assembled, available and searchable. bioRxiv, 2024.https://doi.org/10.1101/2024.03.08.584059

License

GPL-3.0